Prediction Enhancement of Residue Real-Value ... - ACS Publications

MemBrain-Rasa features by an ensemble prediction protocol composed of a statistical machine-learning engine, which is trained in the sequential featur...
0 downloads 6 Views 1MB Size
Subscriber access provided by NEW YORK MED COLL

Article

Prediction enhancement of residue real-value relative accessible surface area in transmembrane helical proteins by solving the output preference problem of machine learning-based predictors Feng Xiao, and Hong-Bin Shen J. Chem. Inf. Model., Just Accepted Manuscript • DOI: 10.1021/acs.jcim.5b00246 • Publication Date (Web): 10 Oct 2015 Downloaded from http://pubs.acs.org on October 12, 2015

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Journal of Chemical Information and Modeling is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

Prediction enhancement of residue real-value relative accessible surface area in transmembrane helical proteins by solving the output preference problem of machine learning-based predictors Feng Xiao and Hong-Bin Shen* Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University, and Key Laboratory of System Control and Information Processing, Ministry of Education of China, Shanghai, 200240, China * Address correspondence to H.B. Shen at [email protected] Tel: +86-21-34205320 Fax: +86-21-34204022

Short title: Membrane protein residue relative accessible surface area prediction Keywords: Alpha-helical membrane proteins; Relative solvent accessibility surface area; Ensemble; Machine learning; Structure template;

1

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 50

ABSTRACT Alpha-helical transmembrane proteins constitute 25% of the entire human proteome space and are difficult targets in high-resolution wet-lab structural studies, calling for accurate computational predictors. We present a novel sequence-based method called MemBrain-Rasa to predict relative solvent accessibility surface area (rASA) from primary sequences. The MemBrain-Rasa features by an ensemble prediction protocol composed by a statistical machine-learning engine, which is trained in the sequential feature space, and a segment template similarity-based engine, which is constructed with solved structures and sequence alignment. We locally constructed a comprehensive database of residue relative solvent accessibility surface area from the solved protein 3D structures in PDB database. It is searched against for segment templates that are expected to be structurally similar to the query sequence’s segments. The segment template-based prediction is then fused with the support vector regression outputs using knowledge rules. Our experiments show that pure machine learning output cannot cover the entire rASA solution space, and will have the serious prediction preference problem due to the relative small size of membrane protein structures that can be used as the training samples. The template-based engine complements this problem very well, resulting in significant improvement of the prediction performance. The MemBrain-Rasa achieves a Pearson correlation coefficient of 0.733 and mean absolute error of 13.593 on the benchmark dataset, which are 26.4% and 26.1% better than existing predictors. MemBrain-Rasa represents a new progress in structure modeling of alpha-helical transmembrane

2

ACS Paragon Plus Environment

Page 3 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

proteins. MemBrain-Rasa is available at: www.csbio.sjtu.edu.cn/bioinf/MemBrain/.

3

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 50

I. INTRODUCTION Membrane proteins (MPs) are mostly located in cell membranes and are encoded by 20-35% of genes 1-3. They have many special and complex physiological functions and take part in various vital activities of cells, including molecular transport, cell recognition and immune response, and signal transduction and energy transfer. MPs are important drug targets and constitute approximately 60% of known drug targets. A transmembrane protein (TMP) is a type of membrane protein (MP) which spans the entire biological membrane to which it is permanently attached. There are two basic types of TMPs: alpha-helical (TMH) and beta-barrels (TMB) according to transmembrane

segments’

classification

of

secondary

structure.

The

membrane-embedded alpha-helical, polytopic proteins constitute the majority of ion channels, transporters, and receptors in living organisms 4. TMH proteins are infamously difficult targets for high-resolution structural studies. Due to the intrinsic structural plasticity associated with many of these proteins, the chance of obtaining crystals suitable for X-ray or electron diffraction studies is small. Although helical membrane proteins present a high degree of experimental difficulty, their conformation is in many ways predictable

5-8

. For

instance, the transmembrane helices must span the hydrophobic layer of membrane or membrane-mimetic detergent micelles, whereas the amphipathic helices or loops are either associated with the head group region or exposed to bulk solvent 9. These conditions effectively speed up the progress of computational model construction. To 4

ACS Paragon Plus Environment

Page 5 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

date, many pseudo-2D space topological structure predictors have been developed, such as the TMH segment prediction tools MemBrain 9, Memsat 10, Memsat-SVM 11, TMHMM

12

, a large-scale benchmark on existing TMH predictors

13

, and the

TMH-TMH contact map prediction 14. Progress in 3D structure prediction for membrane proteins is much slower than for soluble proteins. Only a few membrane-specific protein 3D structure predictors are available, e.g. single chain predictors Membrane-Rosetta

15

and FILM3

16

, and

several easy multi-chain complexes modeling such as BCL::MP-Fold 17. On the other hand, for soluble proteins, there are now many well-established software packages that can be used to predict their 3D structures, e.g., I-TASSER 18, Rosetta 19, TASSER 20

etc. This is due to the rapid increase in number of solved soluble protein 3D

structures in the PDB (Protein Data Bank) 21. According to the recent CASP (Critical Assessment of protein Structure Prediction) competitions, the top ranked predictors are generally the template-based algorithms 22. However, when compared with soluble proteins, MPs’ structures are hard to solve due to the chemical complexity of the biological membrane phospholipid bilayer. For example, in the PDB (released on 2014/05/16), there are in total 103,015 solved structure records, yet only 2,131 proteins are TMPs, of which 1,840 (~1.8%) are alpha-helical and 283 (~0.3%) are beta-barrel membrane proteins. This situation is not expected to improve dramatically in the near future. The lack of high-resolution membrane protein 3D structures and the reality that the membrane protein structure is far complex than previously thought significantly slow down the progress of developing powerful predictors.

5

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 50

Previous studies have demonstrated that accurately predicting the structural features in a pseudo-2D space from the amino acid sequence is an important step to generate a reliable model in the 3D conformational space. In soluble proteins, predictions of residue secondary structure (SS) and relative accessible surface area (rASA) are well studied

23

. Many methods have been developed to predict these

features, such as PSIPRED

24

, SSpro

25

and Spine-X

26

for SS, SABLE

27

for SA

(Solvent Accessibility). In alpha-helical TMPs, there are also several methods to predict SS accurately, such as Memsat TMHMM

12

10

, Memsat-SVM

11

, MemBrain 9, and

for TMH segments prediction. During the past decade, along with many

TMH protein structures being solved, a few methods for predicting membrane protein residue solvent accessibility have also been developed. These can be generally grouped into two classes according to their application area: (1) methods for predicting the residue exposure feature within the membrane, e.g. ASAP 28, TMX 29, and TMexpo 30, and (2) methods that can be applied on the entire membrane protein sequence, both within and outside of the cell membrane regions, e.g. MPRAP

31

.

Table 1 has summarized the features and reported performance of these existing predictors. All the methods for predicting membrane protein residue rASA mentioned above are machine-learning-based predictors. In these approaches, each residue is encoded using a discrete feature vector, which is further fed into a trained statistical machine-learning model for prediction, e.g. support vector machine (SVM) or support vector regression (SVR). It’s worth pointing out that the prediction of rASA on the

6

ACS Paragon Plus Environment

Page 7 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

full sequence is more challenging than those that only apply to residues inside the cell membrane (TMH regions). To the best of our knowledge, the MPRAP is the only predictor that can be applied on the entire membrane protein chain to date. The reason can be due to either more flexibility in the loop region, or a single statistical learning model cannot catch the different features of TMH and loop residues. In this paper, we report

a

more

powerful

whole

sequence-oriented computational approach

(MemBrain-Rasa) to predict real-value rASA of the alpha-helical TMPs. The new algorithm features by a consensus protocol that combines both statistical machine learning-based prediction and segment template-based structural similarity estimates. Benefiting from recent progress in the structural biology, more 3D protein structures are deposited into the PDB, which significantly increases the possibility to find similar segment templates for a given amino acid sequence. In MemBrain-Rasa, the machine-learning engine is achieved by a SVR model which will generate real-value outputs. The input features to SVR include 6 types: position specific scoring matrix (PSSM), predicted evolution rate by rate4site (R4S), predicted residue Z-coordinate, predicted secondary structure (SS) information, representative physical parameters (PP), and sequence length. To fully exploit experimentally solved protein 3D structure information and solve the output preference problem of SVR, we designed a complementary structure template-based estimation engine motivated by the SSPro for soluble proteins constructed a comprehensive

protein

25

. We locally

rASA-orientated database with

their

corresponding amino-acid sequence from all the known 3D structures in PDB. Then

7

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 8 of 50

we use the BLAST software to detect the segment templates in the constructed rASA database for the query membrane protein sequence. The annotated rASA of the homology structural segments will be used to estimate the rASA of their aligned segments on the query sequence. In the case where no structural templates can be found, the final prediction is dependent on the SVR outputs. This new fusion protocol enables MemBrain-Rasa to generate a great improvement on the rASA predictions for the whole chain. On the same benchmark dataset as MPRAP, MemBrain-Rasa achieves a CC of 0.733 and a MAE (mean absolute error) of 13.593, which are much higher than the CC of 0.58 and the MAE of 18.4 achieved by MPRAP. II. MATERIALS AND METHODS 2.1 Membrane protein benchmark dataset In order to fairly evaluate the developed predictor and concurrently remain capable of predicting solvent accessibility of both single- and multiple-chain membrane proteins, we used the same benchmark dataset and the way for calculating the residue rASA values as originally used in MPRAP 31. In this dataset, the sequence identity cutoff was set to 20% and the length cutoff (which specifies the length of coverage) was set to 0.9. Fragments, low-resolution structures and structures with secondary structure or membrane boundary problems were excluded. Finally, this dataset consists of 52 membrane protein complexes composed of 80 chains. This dataset was also divided into five folds in advance to avoid high homology in different folds, where chains from the same superfamily were put in the same fold. All the following experiments of this study are performed on the five folds with a

8

ACS Paragon Plus Environment

Page 9 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

five-fold cross-validation. 2.2 Calculation of rASA The residue solvent accessibility surface area (ASA) is defined as the sum of all atomic surface accessibility in the residue. Several software programs are capable of calculating ASA, such as MSMS 32, Naccess 33, and Dictionary of Protein Secondary Structure (DSSP)

34

. A residue’s ASA is divided by the corresponding standard

accessibilities to generate the relative solvent accessibility surface area (rASA)

35

.

Note that different programs can possibly differ on how to choose the standard, resulting in slight differences in their outputs. For instance, in MSMS, the standard accessibility for each residue is a GLY-X-GLY tripeptide with extended conformation; while in Naccess, an extended ALA-X-ALA tripeptide is used. In this study, the rASA of

each

residue

was

calculated

by

Naccess

(http://www.bioinf.man.ac.uk/naccess), the same as was used in MPRAP

2.1.1 31

. In

Naccess, the probe size 1.4Å mimics water molecules, 2.0 Å mimics CH2-group, and the combination of these two values is used to calculate the rASA of transmembrane proteins. It's worth to note that for making the developed predictor able to predict residue solvent accessibility of both single- and multiple-chain proteins, when in the case of a complex composed by multiple chains, the residues' rASA values are calculated in its complex state, i.e., all chains in the complex are considered simultaneously in their nature complex state. This is different from the single chain calculation mode, where the inter-chain interactions are not considered which will affect the rASA values on the interaction interface.

9

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 50

2.3 Machine-learning-based prediction engine Feature extraction We extracted six types of features from a given amino acid sequence, including position specific scoring matrix (PSSM), evolution rate predicted by rate4site (R4S), Z-coordinate score, predicted secondary structure (SS) information, representative physical parameters (PP), and sequence length. These features can represent both global and local characteristics of the query protein amino acid sequence. (1) Position specific scoring matrix Position specific scoring matrix (PSSM) contains evolutionary information calculated from an aligned set of sequences. In this study, it is generated by PSI-BLAST to search against the UniRef90 database with three iterations and an E-value cutoff of 10-5. The PSSM for each chain is represented by an L × 20 matrix, where L is the sequence length. The raw PSSM scores were scaled to the range [0, 1] using the standardized logistic function as: f (a ) =

1 1 + e −a

(1)

where a is the original score. (2) Evolution rate Rate4Site (R4S) is used to detect conserved amino-acid residues by computing the relative evolutionary rate for sites in the multiple sequence alignment (MSA). According to previous work 36, buried residues will evolve slowly and are considered to be conserved, whereas exposed residues evolve rapidly and are considered to be the opposite. At last, the relative substitution rate was found almost linearly related to the 10

ACS Paragon Plus Environment

Page 11 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

solvent accessibility in membrane protein complexes. In our experiments, the correlation coefficient (CC) between the conserved score alone obtained by Rate4Site and the relative solvent accessibility was 0.44. In addition, conserved score was found the most important in all features mentioned in this paper according to some feature-selection algorithms, i.e. mRMR

37

and stepwise discriminant analysis (SDA)

38

. Considering its importance, the evolution rate score is also incorporated into

MemBrain-Rasa. To restrict the computational time for computing the residue evolution rate, at most the top 50 selected sequences per structure were allowed. If the number of aligned sequences is less than 50, all aligned sequences are included for computation. In Rate4Site, the rate pij of residue i in protein j is first normalized by subtracting the average conservation score p j and dividing by the standard deviation

δ j of the protein j. qij = ( pij − p j ) / δ j

(2)

where qij is the normalized value of pij . The raw score of qij is then further normalized among all the sequences using following equation:

qij' =

qij − min( q j ) max(q j ) − min( q j )

(3)

(3) Z-coordinate Z-coordinate is an important constituent in the field of membrane protein structure prediction and is used to detect the relative position of a residue with respect to the membrane. A decade ago, FILM studied folding in the lipid membrane region, and took the z-axis as perpendicular to the bilayer surface to calculate energy potential 11

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

16

. Membrane-Rosetta

15

and BCL::MP-Fold

17

Page 12 of 50

also used the Z-coordinate to improve

their structure prediction of membrane proteins. Zpred 39 was developed to predict the residue Z-coordinate, and its results represent the relative position of each residue in the membrane. For this paper, MemBrain-Rasa also used the predicted Z-coordinate by Zpred to encode a residue. According to our statistics, almost all of the predicted absolute z-coordinate scores were no more than 25Å, and then the absolute values were normalized by dividing 25, which is used in the final system.

(4) Secondary structure Secondary structure information has proved effective in the field of solvent accessibility and torsion angle prediction of soluble proteins

26, 40, 41

and is also

expected to be useful for this study’s membrane proteins. In this paper, the secondary structure of each residue was predicted by PSIPRED to be one of three classes (coil (C), helix (H) and strand (E)). For each residue, we take its predicted three possibilities directly as the input features.

(5) Representative residue physical parameters Tremendous statistics studies have demonstrated that different amino acids have different properties, resulting in different propensities for forming specific 3D spatial conformations. For instance, the TMH segments in membrane proteins are dominated by hydrophobic amino acids. According to our local tests, we found that the following twelve representative amino-acid physical parameters can improve the system performance and hence were incorporated into the prediction system: a steric parameter, hydrophobicity, volume, polarity, isoelectric point, helix probability, strand

12

ACS Paragon Plus Environment

Page 13 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

probability, average accessible surface area (ASA), charge, acidity, the probability of occurrence, and average mass of twenty common amino acids (supporting information). Note that all of these parameters were normalized to range from 0 to 1.

(6) Sequence length In addition to the features mentioned above, sequence length as a global feature is also used in this study. A one-dimensional vector was used to encode sequence length. In MemBrain-Rasa system, if a query sequence length was less than 900, then its sequence length score was 0.1*⌈ N/100⌉; otherwise, the score was set to 1.

Using sliding window approach to include neighborhood information We then used a sliding window to include more useful residue neighborhood information to improve the prediction accuracy. This makes sense when considering that the status change of neighbor residues is usually continuous. In this study, we used sliding windows to cover neighborhood information in four types of features: (1) PSSM, (2) evolution rate, (3) Z-coordinate, (4) predicted secondary structure information. In our work, we found that the window size of 9 was optimal. Thus, for each residue, we obtained (20 + 1 + 1 + 3) × 9 = 225 encoded components by applying the sliding window on the four aforementioned features. Furthermore, by combining with two other features (representative residue physical parameters and sequence length), each residue was encoded into a 238-D vector (225+12+1). The constructed feature vectors were used to train a SVR model as described next.

Support vector regression Support vector machine (SVM) is a type of supervised learning algorithm, which

13

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 14 of 50

has been proved to be successful in dealing with complex biological data 42. Support vector machines include support vector classification (SVC) and support vector regression (SVR). SVC has been demonstrated to work very well for binary classification, whereas SVR is better for the real value prediction. In our study, SVR was applied since we aim to predict the real-value rASA of alpha-helical TMPs. In SVR, there are three popular kernels that can be used – linear, polynomial, or radial basis functions. In our experiments, the radial basis kernel was used due to its better performance. The parameters in SVR were optimized using grid-search method. SVM-light packages are applied for the model construction, which is available at http://svmlight.joachims.org/.

2.4 The segment template-based prediction In order to improve the prediction of rASA, we not only used machine learning but also developed a segment template-based engine in order to use the experimentally solved 3D structure knowledge for prediction

25

. We constructed a

local residue rASA database using following steps. The database first retrieved all solved structures downloaded from http://www.pdb.org/, then removed the duplicate structures by CD-HIT

43

, resulting in 111,648 structure records. Then the

corresponding rASA index table was constructed by calculating rASA for each structure using the Naccess program. In this way, we have constructed a comprehensive database covering all the recent solved structures. In the prediction step, we developed a homology-segment-based prediction protocol. Given a query sequence, we firstly try to identify the homology sequences

14

ACS Paragon Plus Environment

Page 15 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

from the constructed database using the BLAST program. A homology segment is found if the following five conditions are satisfied: (1) the minimum length of aligned structures is 20, (2) gap is not allowed, (3) the maximum E-value is 10-9, (4) the minimum sequence identity is 70%, and (5) the minimum positive is 75%. To provide a completely fair evaluation, we excluded searched protein chains that are identical to any chain in the benchmark dataset in our following experiments. If segments that meet the above five requirements are identified, they will be used to estimate the residue rASA for the query sequence with the hypothesis that homology sequences will have similar structures. We summed the rASA values of the corresponding positions in the sequences from the structure templates and divided by the total number of templates. The concrete procedure can be seen in Figure 1.

2.5 Hierarchical MemBrain-Rasa consensus prediction system The final system contains both machine-learning SVR-based prediction and the structure similarity-based prediction. To realize a proper combination of the two estimates, we designed the following simple knowledge-guided approach as shown in Figure 2. The input to the final MemBrain-Rasa system is the primary sequence, and the output is the predicted rASA for each residue, which can be summarized in three steps as follows (Figure 3). In the first step, we extract six types of features and use a trained SVR for predictions. In the second step, we use the BLAST program to search against the local rASA database and obtain a prediction for those residues that have homology segments in the rASA database. In the third and final step, we combine both the two predictions together by using the IF THEN knowledge rules. The

15

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 16 of 50

threshold used in the fusion process of Figure 2 is 5 according to the preliminary local experiments.

2.6 Performance evaluation All the results shown in this paper are from five-fold cross-validation. To quantitatively evaluate the real value predictions of relative solvent accessibility prediction, mean absolute error (MAE) and Pearson correlation coefficients (CC) were used. CC is a measure of the linear correlation between predictions and real values. CC value ranges in [-1, 1], where -1 represents total negative correlation, 0 no correlation and 1 total positive correlation. The definition of CC is: L

∑ ( x − x )( y − y ) i

CC =

i

i =1

L  L 2 2 ( x − x ) ∑ i   ∑ ( yi − y )   i =1   i =1 

(4)

where xi and yi represent the i -th residue’s observed and predicted values respectively, x and y represent the corresponding mean value. L represents the total number of residues. The mean absolute error (MAE) measures how close predictions are to their real values and is defined as the average difference between the predicted and observed rASA values of all residues, i.e.

MAE =

1 L ∑ yi − xi L i =1

(5)

III. RESULTS

In the final MemBrain-Rasa system, there are two sub-engines. One is the

16

ACS Paragon Plus Environment

Page 17 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

machine learning SVR-based engine, and the other is the homology segment structure similarity-based engine. Our experimental results show that these two parts complement each other very well, resulting in a significant improvement when compared with the traditional pure machine learning-based predictors.

3.1 Machine learning prediction engine Parameter selection Parameters are one of the key factors of a machine-learning algorithm, and different parameters can result in very different results. In this part, there are a total of three parameters: window length for encoding the features and the two parameters of “C” and “g” in the SVR algorithm. For the window length, different sizes of a symmetric window were tested and we found 9 to be optimal and used in this study, which is the same as that in MPRAP

31

. In the SVR algorithm, “C” represents the

trade-off between training error and margin, and “g” represents parameter gamma in the RBF kernel function. These two parameters were optimized through a grid search approach with five-fold cross-validation, where the parameters corresponding to the best prediction results are used. In the final version of SVR model, we selected radial basis kernel function (RBF) with g=0.04 and C=18.

Feature fusion improves the prediction accuracy A basic rule for the statistical machine-learning algorithm is “good input, good output.” In order to make the SVR learn effective rules from the dataset, the input features to SVR must be useful. However, not all the features that can be calculated are useful for the residue rASA prediction. To test the six features discussed above,

17

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 50

we have tested both independent features and their combinations. Table 2 shows the results. All the four independent features (PSSM, R4S, SS, and Zpred) contain some useful information for predicting rASA by themselves as is shown in Table 2. Among these independent inputs, PSSM achieves the best overall results (MAE=23.00 and CC=0.485). This result suggests that PSSM is important in rASA prediction, possibly due to the existence of some residue evolutionary knowledge. R4S is tested as the second-best feature for predicting the residue rASA values. The success of the R4S feature is also due to its strong correlation with residue evolution rates and is consistent with previous research 36. When combining multiple features, we find that the best CC values always accompany the least MAE, indicating that the two evaluation matrices are consistent, as was also shown by a previous study that the CC values are almost linearly related to the MAE values

28

. Since our independent feature test has shown that PSSM and

R4S have the best results on rASA prediction, we added these two features together first. The sequence encoding scheme of the combination of “PSSM+R4S” increases the prediction accuracy to CC=0.525, which is a significant improvement compared to the CC=0.485 (PSSM alone). Afterwards, the encoding scheme “PSSM+R4S+Zpred” which adds the Zpred information gives another obvious improvement. Compared to “PSSM+R4S”, the MAE value improves by almost 4% and the CC value reaches 0.58 for the new input. This is likely due to the fact that the Z-coordinate predicted by Zpred is a different point of view to the target residue from the PSSM and R4S and

18

ACS Paragon Plus Environment

Page 19 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

contains the information that PSSM and R4S don’t have, resulting in a much better prediction model. The performance of the “PSSM+R4S+Zpred” model is very comparable to those reported in the literature of MPRAP (MAE=18.4 and CC=0.58). To further test whether other features also contribute, we continued to add SS, sequence length, and PP in turn to the feature set. Our results show that we obtain the best overall result when combining all the independent features together, with CC= 0.599 and MAE=18.013. Through this stepwise feature test process, we finally selected the encoding scheme of “PSSM+R4S+Zpred+SS+length+PP” to generate a SVR prediction model.

Neither dimension reduction nor two-stage SVR model considerably help to improve the performance In order to improve prediction performance, in addition to adding new features, we have made many other attempts. We first considered that there could be redundant components in the feature set; thus we applied two feature-selection methods which have been proven effective to solve classification problems in the field of machine learning and bioinformatics

38, 44

. The two tested feature-selection methods are the

minimum redundancy and maximum relevance feature-selection (mRMR) method, which is based on mutual information calculations

37

, and the stepwise discriminant

analysis (SDA) method. They use statistical methods to exclude redundant features. However, the results are not as expected, whereby the two feature selection methods didn’t result in obvious improvement. This may indicate that the distribution of the information for the prediction is not in a compact space, making the problem more

19

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 20 of 50

difficult. Previous studies have also indicated that a two-stage model performed well for predicting solvent accessibility in water-soluble proteins

41, 45

. We also tried to use a

two-stage SVR model rather than a simple SVR on current problem to see whether we can improve the performance. However, in our work, a two-stage SVR model is found to only improve the mean absolute error (MAE) but decrease the CC index simultaneously. The result difference on membrane and water-soluble proteins may be due to smaller sample size in membrane proteins compared to the globular proteins.

3.2 Structure segment template-based prediction engine In the segment structural similarity prediction part, the key idea is to find the structure templates and use them to help predict the rASA of the query sequence. There are two important parameters during this homology-segment-based prediction process. The first is how to judge two sequences are close enough when using the BLAST tool to search the query sequence against the local rASA database; the other is how to remove noise in the alignment by selecting a proper segment length. In order to get the optimal combination of above two parameters, we used a modified grid-search strategy. First we set that a valid segment length needed to be at least 20 (aligned sequences whose lengths are shorter than 20 were excluded from further consideration); next we determined which sequence similarity will generate the best result by searching from 0 to 100% in steps of 10%. The results of MemBrain-Rasa at different sequence identity cutoff thresholds are shown in Figure 4A. The difference between the two sets of data in this figure is the

20

ACS Paragon Plus Environment

Page 21 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

inclusion of identical segments. Obviously, the results of including the identical segments are considerably better than those where identical segments are excluded. To get a fair and strict result, we mainly evaluate the performance when identical segments are excluded. It is clear that when we count only those searched segments that are of 100% sequence identity to the segments on the query sequence, the final results are solely from the SVR predictions, which are CC=0.599 (Table 2). It’s interesting to observe that the prediction results are very stable on tested thresholds. The reason could be that we have set a strict homology segment screening condition (the minimum length of aligned segments should be at least 20 amino acids long, gap is not allowed, and the minimum E-value is 10-9), which gives us relatively stable searched result sets on all the thresholds. Finally, in the MemBrain-Rasa system, the minimum sequence identity for screening a homology segment is set to 70%. For the second parameter of the homology segment length, we fixed the minimum sequence identity to be 70% and then searched the sequence length from 0 to 100 in steps of 10. We found that prediction performance changes very little when the sequence length is lower than 60 (Figure 4B). Finally, we decided that a sequence length of 20 would be used in the MemBrain-Rasa.

3.3 Consensus model helps to solve the preference prediction problem of SVR The SVR algorithm is a typical statistical machine-learning-based predictor, which learns the regression parameters from a pre-collected dataset. The final trained model is greatly affected by the dataset distribution. Considering the general problem of incompleteness in the pre-collected training dataset, the trained SVR model will

21

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 22 of 50

not be a perfect predictor. Thus, the SVR-based predictor usually has some preferences to make a prediction in specific regions due to having learned more rules from them in the training dataset. The incompleteness of a training dataset is more serious in this paper due to the relatively small number of solved membrane protein 3D structures. Figure 5 illustrates the distribution of ten subsets of the benchmark residue rASA value, with the pure SVR-based predictions and the consensus MemBrain-Rasa predictions are also listed. As can be seen, the distribution of the benchmark residue rASA is not balanced, where 8,300 residues lie in the group of (0, 10], whereas only ~1,000 residues are located in the three groups of [70, 80], [80, 90], and [90, 100]. The effect of this imbalanced dataset distribution on the trained SVR predictor is obvious, resulting in bad performance of the trained SVR on those three groups. Especially in the groups [80, 90] and [90, 100], where almost no predictions are made. In the proposed MemBrain-Rasa predictor, the combination of structure template-based engine with the machine learning prediction is capable of resolving the effects of imbalanced datasets on the machine learning predictors. As can be seen from Figure 5, the distribution of MemBrain-Rasa’s outputs is much closer to the benchmark as compared to the pure SVR method. This significant improvement implies that we can now predict residues that lie in region of [70, 100] more accurately, especially in region [70, 80]. The reason for this improvement is that the homology segment-based prediction does not need the training process. In order to further demonstrate the superiority of the proposed MemBrain-Rasa

22

ACS Paragon Plus Environment

Page 23 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

protocol over the pure SVR-based predictor, Figure 6 illustrates the CC and MAE distribution between the prediction and benchmark on 80 alpha-helical TMPs in the benchmark dataset. We can see that when only using SVR model, 59 (59/80=73.75%) chains have a predicted CC over 0.5. However, no chain has a CC greater than 0.8. These results suggest that SVR-based predictor cannot generate very accurate predictions. In the developed MemBrain-Rasa predictor, we combined the homology segment structural similarity-based predictor with the SVR-based predictor. Our results show that there are 49 proteins (with a total of 13,306 residues) that can find some homology segments in the local constructed rASA database out of all 80 proteins (21,418 residues). Among the 13,306 residues, 8,091 residues need to be modified from the original SVR predictions to the segment template-based predictions because the absolute difference between these two predictions is more than 5 (see combination rule in the Figure 2). The great difference between SVR’s output and MemBrain-Rasa’s is that there are 27 (27/80=33.8%) chains with CC values over 0.8 in MemBrain-Rasa but none in SVR. The MAE distribution in Figure 6(B) is very similar to the CC distribution. There are 27 chains with MAEThreshold ) Then Else

rASA=Hom rASA=ML

Figure 2. Knowledge-guided approach to hierarchically combine machine learning and structure similarity-based predictions. Threshold is used to denote the deviation from the two prediction engines and 5 is used in this study.

44

ACS Paragon Plus Environment

Page 45 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 V Q L A H H F S E P E I T L I I F G V MA G V I G T I L L I S Y G I R R L I K K

1 1.1 0.9 1 0.8

0.9

0.7

0.8

0.6

0.7

0.5

0.6

0.4

0.5

0.3

0.4

0.2

0.3 0.2

0.1

0.1

0 V Q L A H H F S E P E I T L I I F G V MA G V I G T I L L I S Y G I R R L I K K

0 V Q L A H H F S E P E I T L I I F G V MA G V I G T I L L I S Y G I R R L I K K

Figure 3. Flowchart of the three-step approach in MemBrain-Rasa. (a) The procedure of machine learning predictor. (b) The procedure of segment structure similarity-based predictor. (c) The final rASA predictor by combining both machine learning and structure similarity-based predictors.

45

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 46 of 50

Figure 4. (A) CC values between the observed and predicted rASA with different minimum percentages of sequence identity. (B) CC values between the observed and predicted rASA with different homology segment lengths.

46

ACS Paragon Plus Environment

Page 47 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

Figure 5. Distribution of SVR and MemBrain-Rasa predictions, as well as the residue benchmark rASA value.

47

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 48 of 50

Figure 6. (A) The distributions of CC of rASA predictions for 80 sequences in the benchmark dataset. (B) The distributions of MAE of rASA predictions for 80 sequences in the benchmark dataset.

48

ACS Paragon Plus Environment

Page 49 of 50

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Journal of Chemical Information and Modeling

Figure 7. (A) The rASA predictions of 2bs2:C using SVR model (MAE=18.39, CC=0.63) (B) The rASA predictions of 2bs2:C using the combination of SVR models and structure similarity-based method (MAE=4.93, CC=0.95).

49

ACS Paragon Plus Environment

Journal of Chemical Information and Modeling

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 50 of 50

For Table of Contents use only

Prediction enhancement of residue real-value relative accessible surface area in transmembrane helical proteins by solving the output preference problem of machine learning-based predictors Feng Xiao and Hong-Bin Shen

50

ACS Paragon Plus Environment